English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

Are foundation models secure from malicious actors? In this work, we focus onthe image input to a vision-language model (VLM). We discover image hijacks,adversarial images that control generative models at runtime. We introduceBehaviour Matching, a general method for creating image hijacks, and we use itto explore three types of attacks. Specific string attacks generate arbitraryoutput of the adversary's choice. Leak context attacks leak information fromthe context window into the output. Jailbreak attacks circumvent a model'ssafety training. We study these attacks against LLaVA, a state-of-the-art VLMbased on CLIP and LLaMA-2, and find that all our attack types have above a 90%success rate. Moreover, our attacks are automated and require only small imageperturbations. These findings raise serious concerns about the security offoundation models. If image hijacks are as difficult to defend against asadversarial examples in CIFAR-10, then it might be many years before a solutionis found -- if it even exists.

Image Hijacks: Adversarial Images can Control Generative Models at Runtime